World Models: Objective Dimensionality Dictates Learned Task Structure

This paper investigates how much of a task's underlying structure a learned world model actually represents. Contrary to assumptions that model quality is inherent, the authors demonstrate that the critical factor is the dimensionality of the objective function used for training. When a model is trained with a simple scalar value signal, it only captures a limited, one-dimensional aspect of a potentially multi-dimensional task closure. The study used a DreamerV3 stack in a controlled environment to measure this effect directly. They found that increasing the objective's dimensionality from one to four directly led to the installation of a corresponding number of predictive directions within the model. This suggests that the objective dictates what latent representations are formed, particularly in scenarios where simpler training signals might not suffice.